CN-121998844-A - Microscopic image self-adaptive deconvolution method, device, storage medium and imaging system

CN121998844ACN 121998844 ACN121998844 ACN 121998844ACN-121998844-A

Abstract

The invention discloses a microscopic image self-adaptive deconvolution method, a device, a storage medium and an imaging system, wherein in the method, two image pairs of each sampling area in a plurality of different sampling areas of a target sample are obtained as samples, and each sample forms a training set; the method comprises the steps of training an adaptive deconvolution model comprising main branches and secondary branches through a training set, inputting an image to be deconvoluted into the main branches in the trained adaptive deconvolution model to obtain a corresponding deconvolution image, and using equipment and a storage medium for realizing a method process, wherein the system comprises an optical imaging system for acquiring the image, and a control and data processing system for controlling the optical imaging system and realizing the method process. According to the self-adaptive deconvolution method, the deconvolution network, the optical point spread function network and the loss function are carefully designed, so that deep characteristic representation of training data and microscopic imaging optical degradation representation contained in the data are automatically mined, and the purpose of self-adaptive deconvolution is achieved.

Inventors

WANG XUANKAI
WANG ZONGFA
WANG YUZHUO

Assignees

安徽省太微量子科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (10)

1. The self-adaptive deconvolution method of the microscopic image is characterized by comprising the following steps of: Acquiring images of a plurality of different sampling areas of a target sample, and constructing a plurality of image pairs based on the acquired images, wherein each image pair comprises two images of the same sampling area, and the sampling areas of the different image pairs are different; Training an adaptive deconvolution model through a training set, wherein the adaptive deconvolution model comprises a deconvolution network and an optical point spread function network, and the training process is as follows: The method comprises the steps of inputting an original image of each sample in a training set into a deconvolution network, obtaining deconvolution images of the original image of each sample through the deconvolution network, generating prior fuzzy kernels through an optical point spread function network, convolving the deconvolution images of the original image of each sample with corresponding prior fuzzy kernels to obtain optically degraded images, calculating error difference values through a loss function based on target images of each sample, corresponding deconvolution images and optically degraded images, obtaining gradients of the error difference values, and reversely propagating the gradients to update parameters of the deconvolution network and the point spread function network; and obtaining an image to be deconvoluted of a target area of the target sample, inputting the image to be deconvoluted into a deconvolution network in the trained self-adaptive deconvolution model, and outputting a corresponding deconvolution image by the deconvolution network.
2. The adaptive deconvolution method of microscopic images according to claim 1, wherein an image of each sampling area of the target sample under high-power illumination and low-power illumination is obtained, wherein the image under high-power illumination is an image with higher signal-to-noise ratio of each sampling area, the image under low-power illumination is an image with lower signal-to-noise ratio of each sampling area, the image with higher signal-to-noise ratio of each sampling area is taken as a target image, the image with lower signal-to-noise ratio is taken as an original image, an image pair is formed by the target image and the original image of each sampling area, each image is taken as a sample respectively, and a supervised modal training set is formed by each sample; And then, performing supervised modal training on the self-adaptive deconvolution model through a supervised modal training set.
3. The self-adaptive deconvolution method of microscopic images according to claim 1, for a training set in a self-supervision training mode, characterized in that images of a plurality of different sampling areas of a target sample under low-power illumination are acquired, data amplification is carried out on the images of each sampling area to obtain two images with different signal-to-noise ratios of each sampling area, one image is randomly selected from the two images of each sampling area to serve as an original image, the other image is selected as a target image, an image pair is formed by the target image and the original image in each sampling area, each image is respectively taken as a sample, and the self-supervision mode training set is formed by each sample; And then, performing self-supervision mode training on the self-adaptive deconvolution model through a self-supervision mode training set.
4. The method of claim 1, wherein the deconvolution network in the adaptive deconvolution model is any convolutional neural network or is any Vision Transformer.
5. The method of claim 1, wherein the optical point spread function network in the adaptive deconvolution model is any convolutional neural network or is any VisionTransformer.
6. The method of claim 1, wherein the loss function of the adaptive deconvolution model during training comprises a fidelity loss term, a hessian loss term, and a perceptual loss term.
7. The adaptive deconvolution method of microscopic image according to claim 1, wherein when training is completed until the loss error value calculated by the loss function converges, the training is ended, and a trained adaptive deconvolution model is obtained.
8. An electronic device comprising a processor and a memory, wherein program instructions in the memory, when read and executed by the processor, perform the adaptive deconvolution method of a microimage as in any of claims 1-7.
9. A storage medium storing program instructions which, when read and executed, perform the method of adaptive deconvolution of microscopic images according to any of claims 1-7.
10. An adaptive deconvolution microscopic imaging system, comprising: An optical imaging system (100) for optically imaging a target sample and acquiring images of regions on the target sample; a control and data processing system (200) for controlling said optical imaging system (100), for acquiring images obtained by the optical imaging system (100), and for implementing an adaptive deconvolution method of microscopic images according to any of claims 1-7.

Description

Microscopic image self-adaptive deconvolution method, device, storage medium and imaging system Technical Field The invention relates to the field of microscopic image processing methods, in particular to a microscopic image self-adaptive deconvolution method, device, storage medium and imaging system. Background Optical microscopic imaging is widely applied to research of life science, material science and industrial detection by the characteristics of mildness, rapidness, simple sample preparation, convenient use and the like. However, due to the existence of the optical diffraction limit, the imaging resolution of the common optical microscope is limited to 200 nm or more, and a finer and key substance structure cannot be observed. In recent years, with the continuous improvement and promotion of computer hardware, computational optical imaging has greatly developed, and through algorithm post-processing, the resolution of microscopic images can be improved by 1.5-2 times, which has great significance for analyzing microscopic subcellular interactions and microscopic defects of materials and the like. The existing deconvolution algorithm for improving the resolution of the microscopic image is mainly divided into two main categories, namely (1) a deconvolution method based on an analysis model and (2) a deconvolution method based on deep learning. Deconvolution methods based on analytical models, such as Richarson-Lucy (RL) deconvolution, fast iterative soft-threshold deconvolution (FISTA), and spark deconvolution, first artificially impose some noise reduction and regularization constraints (such as continuity, sparsity, symmetry, etc. constraints) on the optical imaging model, and then improve image resolution by continually scaling and inverting iterations. The algorithm is easy to be influenced by parameter setting, artificial model constraint, image noise level and iteration times, and generalization and robustness are to be improved. The deconvolution algorithm based on deep learning, such as DFCAN, meta-rLLS-VSIM and the like, avoids introducing artificial physical model constraint and experience parameter setting through the direct mapping relation between end-to-end learning data, and is superior to the analysis model algorithm in generalization and robust performance. However, deep learning algorithms require a large amount of paired training data to be collected in advance to build the network training, which is often very expensive and time consuming. The ZS-DeconvNet algorithm adopts a self-supervision mode to construct deconvolution network training, so that the problems are relieved to a certain extent. But ZS-DeconvNet requires prior provision of the neural network with prior information of the optical imaging system point spread function PSF in order to incorporate the physical prior of the optical imaging process into the network training. The PSF acquisition is usually accompanied by random noise interference and systematic aberration effects, which in turn can also affect the end effect of ZS-DeconvNet, producing a series of deconvolution artifacts such as ringing effects. Disclosure of Invention The invention provides a microscopic image self-adaptive deconvolution method, device, storage medium and imaging system, which are used for solving the problems of deconvolution artifact and deconvolution instability caused by unmatched point spread function prior in a ZS-DeconvNet algorithm for microscopic image deconvolution in the prior art. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: the self-adaptive deconvolution method of the microscopic image comprises the following steps: Acquiring images of a plurality of different sampling areas of a target sample, and constructing a plurality of image pairs based on the acquired images, wherein each image pair comprises two images of the same sampling area, and the sampling areas of the different image pairs are different; Training an adaptive deconvolution model through a training set, wherein the adaptive deconvolution model comprises a deconvolution network and an optical point spread function network, and the training process is as follows: The method comprises the steps of inputting an original image of each sample in a training set into a deconvolution network, obtaining deconvolution images of the original image of each sample through the deconvolution network, generating prior fuzzy kernels through an optical point spread function network, convolving the deconvolution images of the original image of each sample with corresponding prior fuzzy kernels to obtain optically degraded images, calculating error difference values through a loss function based on target images of each sample, corresponding deconvolution images and optically degraded images, obtaining gradients of the error difference values, and reversely propagating the gradients to update parameters of the deconvolution netw